Abstract
Background
To develop and validate a machine learning-based predictive model for identifying high axillary nodal burden (≥ pN2) in SLNB-positive breast cancer patients, integrating preoperative, intraoperative, and pathological features to support clinical decision-making when ALND is omitted.
Methods
A retrospective study was conducted at a single institution between 2020 and 2024, enrolling 956 breast cancer patients who underwent both SLNB and ALND. Clinicopathological variables (preoperative imaging, intraoperative SLN findings, postoperative pathology) were collected. The cohort was randomly split into a training set (70%, n = 670) and a validation set (30%, n = 286). Features were selected via univariate analysis (p < 0.05), LASSO regression (λ1se criterion), and the Boruta algorithm. Nine machine learning models were trained and evaluated using AUC, accuracy, sensitivity, specificity, and F1-score. SHAP analysis was used to interpret the optimal model, and a clinical nomogram was developed based on logistic regression.
Results
The prevalence of high nodal burden (≥ pN2) was 15.3% (146/956). Seven key predictors were identified: Neurological invasion、Lymph vascular invasion、Imaging abnormal nodes、Lymph nodes size、Tumor size、Negative SLN、Positive SLN. Logistic Regression (validation AUC: 0.850, accuracy: 0.857, F1-score: 0.602) and Support Vector Machine (validation AUC: 0.857, accuracy: 0.853, F1-score: 0.596) demonstrated the most stable performance across datasets. A nomogram was developed for intuitive clinical use.
Conclusion
The proposed machine learning model accurately predicts high nodal burden in SLNB-positive breast cancer patients, facilitating individualized adjuvant therapy planning and avoiding unnecessary ALND.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12957-026-04194-w.
Keywords: Breast cancer, ALND, Machine learning, Nomogram
Introduction
Globally, breast cancer exhibits the highest incidence among women, accounting for 25% of all cancer cases and 15.5% of cancer-related deaths [1], posing a serious threat to women’s physical and mental health. Sentinel lymph node biopsy (SLNB) is currently the standard procedure for axillary staging in patients with clinical-node-negative (cN0) breast cancer [2]. Although axillary lymph node dissection (ALND) was routinely performed for patients with positive SLNs, current evidence demonstrates that ALND can be safely omitted in patients with 1 or 2 SLN metastases. Randomized trials such as ACOSOG Z0011 [3] and AMAROS [4] confirmed that in clinical T1–2 breast cancer patients with 1 or 2 SLN positive, there was no significant difference between axillary radiotherapy alone and ALND in terms of 10-year axillary recurrence rate, DFS and OS. Negative outcomes of ALND, such as arm lymphedema (affecting up to 25% of patients), numbness, and restricted shoulder mobility, are well documented [5]. Arm morbidity is considerably worse after ALND than after SLNB only. However, omitting ALND also means that complete axillary staging information cannot be obtained, especially if there is high nodal burden (≥ 4 positive lymph nodes) [2]. High axillary nodal burden is strongly linked to recurrence risk in breast cancer patients and serves as a critical determinant for selecting postoperative adjuvant therapy [6]. Clinical trials have established that patients with > 4 metastatic lymph nodes following ALND derive benefit from regional nodal irradiation (RNI); however, the benefit for those with 1–3 positive nodes remain controversial [7]. The recent SUPREMO trial found that PMRT did not improve overall survival in women with 1–3 positive nodes, though it did reduce regional recurrence, highlighting the ongoing complexity of adjuvant treatment decisions in this group. For patients with estrogen receptor (ER)-positive, HER2-negative breast cancer, selection of certain systemic therapies depends on high nodal burden. The recently reported MONARCH E trial demonstrated that adding 2 years of CDK4/6 inhibitor adjuvant therapy to endocrine treatment significantly improves invasive disease-free survival in high-risk early-stage ER-positive breast cancer (defined as ≥ 4 metastatic axillary lymph nodes, or 1–3 metastatic lymph nodes with tumor size ≥ 5 cm or histological grade 3) [8]. The MINDACT trial further confirmed that axillary nodal burden critically informs systemic treatment strategies [9]. Consequently, nodal metastatic burden substantially influences adjuvant therapy decisions, and omitting ALND could result in undertreatment.
Consequently, to advance precision therapy, clinically applicable tools must be developed to predict high nodal burden when ALND is omitted. Although studies have attempted predictive models for non-sentinel lymph node metastasis or high nodal burden [10], most rely on logistic regression—constrained by linear assumptions and limited predictive performance. Machine learning is increasingly applied in biomedicine, showing strong potential in clinical diagnosis, precision therapy, and health monitoring [8]. This study focused on SLN-positive patients to develop a predictive model for high axillary nodal burden assessment post-surgery, addressing the information gap resulting from ALND omission. We integrated preoperative clinical characteristics, intraoperative SLN indicators, and tumor biological features to construct models using diverse algorithms (including traditional logistic regression and machine learning methods), evaluated performance metrics, identified optimal algorithms, and enhanced interpretability via SHAP analysis. A clinical nomogram was subsequently developed for visualization. This model aims to provide precise guidance for postoperative adjuvant therapy stratification in breast cancer, supporting axillary management in advancing toward individualization and de-escalation strategies.
Materials and methods
Study population
This retrospective analysis reviewed 1,100 breast cancer patients treated with SLNB and subsequent ALND at Jiangsu Provincial Hospital from 2020 to 2024. The study focused on patients with pT1-2 tumors and 1–3 positive SLNs.We excluded those who received neoadjuvant chemotherapy (NAC) due to its potential to alter tumor staging, as well as patients with pathologically negative SLNs, as their low nodal burden status was already confirmed and thus not relevant to the study aim. The final cohort comprised 956 patients, who were randomly allocated to a training set (70%, n = 670) or a validation set (30%, n = 286).
Data collection
Clinical-pathological variables were extracted from medical records, including age, menopausal status, axillary lymph node maximal short-axis diameter, imaging findings, tumor location/multifocality, histological grade, tumor size, lymphovascular invasion (LVI), perineural invasion, number of positive/negative SLNs, hormone receptor (ER/PR/AR) status, HER2 status, and Ki-67 index.
Classification
According to immunohistochemical (IHC) analysis, estrogen receptor (ER) status was defined as follows: negative for < 1% nuclear staining, weakly positive for 1% to 10% nuclear staining, and positive for > 10% nuclear staining [11]. Progesterone receptor (PR) status was classified as negative (< 1% staining), weakly positive (1% to 20% staining), or positive (> 20% staining). For human epidermal growth factor receptor-2 (HER-2), an IHC score of 0 was considered negative. Low HER-2 expression was defined as an IHC score of 1 + or an IHC score of 2 + accompanied by a negative fluorescence in situ hybridization (FISH) result. HER-2 positivity was defined as an IHC score of 3 + or an IHC score of 2 + with a positive FISH result [12]. Ki-67 expression levels were categorized into two groups: low (Ki-67 ≤ 14%) and high (Ki-67 > 14%) [13]. Based on the IHC profiles, all patients were classified into one of four molecular subtypes: Luminal A, Luminal B, triple-negative breast cancer (TNBC), or HER-2 overexpression. For the AR variable, immunohistochemical (IHC) staining was considered positive if > 10% of tumor cell nuclei showed staining, and negative otherwise [14]. Immunohistochemical staining for CK5/6 was assessed as positive when ≥ 10% of tumor cells showed any perceptible cytoplasmic staining. Negative staining was defined as the absence of staining or staining in < 10% of tumor cells [15]. p53 positivity, indicative of abnormality, was defined as either overexpression (> 70% strong nuclear staining) or complete loss (< 1% staining). All other patterns were deemed negative [16]. Lymphocyte Percentage was derived from serum biochemistry. Suspicious lymph nodes on ultrasound were graded for malignancy risk: 4a (2–10%), 4b (10–50%), or 4c (50–94%). The presence of microcalcifications in invasive carcinoma was confirmed by pathologic observation of characteristic dark-blue, granular deposits within the tumor or adjacent DCIS on H&E-stained sections.
Machine learning models and evaluation
Nine machine learning (ML) models were employed to predict the risk of high nodal burden: Random Forest (RF), Logistic Regression (LR), Gradient Boosting Machine (GBM), Neural Network (NN), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM), and Support Vector Machine (SVM). Several commonly used evaluation metrics [17]were applied, including the area under the Receiver Operating Characteristic (ROC) curve (AUC), sensitivity, specificity, accuracy, and F1 score, to assess the reliability of these models.
SHAP interpretability analysis
SHAP (SHapley Additive exPlanations) was used to enhance model interpretability. Derived from cooperative game theory Shapley values, it attributes each feature’s contribution to individual predictions additively, offering granular model behavior insights—particularly for complex machine learning models [18]. For the SVM model, the shapviz package computed/visualized SHAP values (per https://github.com/slundberg/shap) to identify high nodal burden predictors, generating summary (bar), beeswarm, and waterfall plots; features with larger absolute SHAP values had greater predictive impact. SHAP summary/waterfall plots were also applied to two representative patients (with/without high nodal burden) to illustrate individual-level interpretability, clarifying how specific features drove each patient’s predicted outcome.
Statistical analysis
Statistical analyses (data preprocessing, model development/assessment) were conducted in R 4.4.1. Missing values were addressed via multiple imputation. Normally distributed continuous variables were presented as mean ± SD (x̄ ± SD) and compared with Student’s t-test; categorical variables as n (%) and compared with Chi-square or Fisher’s exact test (as appropriate). Statistical significance was set at two-sided p < 0.05. Models (based on identified risk factors) were internally validated via the validation cohort. ROC curves calculated AUC (AUC > 0.7 = acceptable discrimination) to assess model performance; calibration curves and DCA evaluated model calibration and clinical utility, respectively.
Study outcome
The primary outcome of this study was the pathological presence of high axillary nodal burden, defined as ≥ 4 metastatic axillary lymph nodes (stage pN2 or higher) confirmed by final histopathology after ALND. This outcome was dichotomized (yes/no) for model prediction. The intended application of this predictive model is postoperative risk stratification for patients in whom ALND is omitted. By integrating readily available preoperative, intraoperative, and postoperative pathological data, the model serves as a surrogate tool for comprehensive recurrence risk assessment. It aims to identify, among patients with 1–3 positive SLNs who forego ALND, those with a high likelihood of harboring ≥ pN2 disease. This information is critical for tailoring adjuvant therapy (such as systemic treatment intensification and radiotherapy planning), thereby mitigating the risk of undertreatment resulting from incomplete axillary staging.
Result
Patient characteristics
The overall design and workflow of this study are summarized and illustrated in Fig. 1. In this study, we performed a descriptive analysis of pre-and post-operative clinical information on 956 breast cancer patients who underwent ALND. According to pathological lymph node metastasis status, the patients were divided into low lymph node metastasis burden group (pN1, n = 810) and high lymph node metastasis burden group (≥ pN2, n = 146). It can be seen that among SLNB positive ALND patients, only 15.3% (146/956) eventually experienced a transition from low to high lymph node burden. Thus, ALND can be considered overtreatment for the remaining 84.7% of patients. shows baseline characteristics for both groups, with patients in the high-burden group more likely to have radiographically abnormal lymph node ≥ 5 mm (54.8% vs. 44.8%; P = 0.033), BI-RADS 4 B-type suspicious lymph node (28.8% vs. 20.1%; P = 0.016), and a significantly higher proportion of tumors located in the non-upper lateral quadrant (UOQ) (58.9% vs. 47.0%; P = 0.011). In terms of molecular pathological features, the high expression rate of androgen receptor (AR) (≥ 95%) in the high-load group was significantly increased (57.5% vs. 40.1%; P < 0.001), while the positive rate of cytokeratin CK5/6 was significantly decreased (8.2% vs. 15.4%; P = 0.031). Pathological analysis further revealed that the incidence of lymphatic vessel invasion (67.8% vs. 30.2%; P < 0.001) and nerve invasion (36.3% vs. 9.6%; P < 0.001) was significantly increased in the high load group. Sentinel lymph node (SLN) analysis showed that the median positive SLN was significantly higher in the high-burden group (2.00 vs. 1.00; P < 0.001), while the median negative SLN was significantly lower (1.50 vs. 3.00; P < 0.001). Table S1 shows the comparison of baseline characteristics between training set and validation set. Specifically, except PR (PR ≥ 95%: 24.0% vs. 31.1%, P = 0.028), there was no statistically significant difference between groups for all variables (all P > 0.05), indicating that a high degree of equilibrium was maintained when randomly dividing training set and validation set, and the two data sets were homologous and comparable.
Fig. 1.
Flow chart of the study
Cohort division and feature selection
We analyzed the correlation between clinicopathological features and axillary lymph node (ALN) status. The dataset was randomly stratified into a training cohort (70%, n = 670) and a validation cohort (30%, n = 286). When screening variables, we first conducted a one-way analysis of each variable, and the results are shown in Table 1. To further refine the feature selection process, we combined LASSO regression with the Boruta algorithm. As detailed in.
Table 1.
Baseline characteristics of the whole dataset
| Characteristics | Total (n = 956) | Low nodal burden, (n = 810) | High nodal burden, (n = 146) | P value |
|---|---|---|---|---|
| Age, median (IQR) | 54.00 [45.00, 61.00] | 54.00 [46.00, 61.00] | 52.00 [45.00, 61.00] | 0.368 |
| Menstrual (%) | ||||
| Premenopausal | 432 (45.2) | 368 (45.4) | 64 (43.8) | 0.79 |
| Postmenopausal | 524 (54.8) | 442 (54.6) | 82 (56.2) | |
| Lymphocyte (%) | ||||
| < 30% | 493 (51.6) | 416 (51.4) | 77 (52.7) | 0.828 |
| ≥ 30% | 463 (48.4) | 394 (48.6) | 69 (47.3) | |
| Lymph nodes size (%) | ||||
| < 5 mm | 513 (53.7) | 447 (55.2) | 66 (45.2) | 0.033 |
| ≥ 5 mm | 443 (46.3) | 363 (44.8) | 80 (54.8) | |
| Imaging abnormal nodes (%) | ||||
| 4 A | 515 (53.9) | 436 (53.8) | 79 (54.1) | 0.016 |
| 4B | 205 (21.4) | 163 (20.1) | 42 (28.8) | |
| 4 C | 236 (24.7) | 211 (26.0) | 25 (17.1) | |
| Tumor location (%) | ||||
| other | 467 (48.8) | 381 (47.0) | 86 (58.9) | 0.011 |
| UOQ | 489 (51.2) | 429 (53.0) | 60 (41.1) | |
| ER (%) | ||||
| Negative | 180(18.8) | 162 (20.0) | 18 (12.3) | 0.092 |
| Low | 44(4.60) | 37 (4.6) | 7 (4.8) | |
| Positive | 732(76.5) | 611 (75.4) | 121 (82.9) | |
| PR (%) | ||||
| Negative | 251(26.3) | 226 (27.9) | 25 (17.1) | 0.024 |
| Low | 79(8.3) | 66 (8.1) | 13 (8.9) | |
| Positive | 626(65.4) | 518 (64.0) | 108 (74.0) | |
| Ki67 (%) | ||||
| < 14% | 108(11.3) | 90 (11.1) | 18 (12.3) | 0.775 |
| ≥ 14% | 848(88.7) | 720 (88.9) | 128 (87.7) | |
| HER2 (%) | ||||
| Negative | 751 (78.6) | 642 (79.3) | 109 (74.7) | 0.255 |
| Positive | 205 (21.4) | 168 (20.7) | 37 (25.3) | |
| P53 (%) | ||||
| Negative | 689 (72.1) | 588 (72.6) | 101 (69.2) | 0.456 |
| Positive | 267 (27.9) | 222 (27.4) | 45 (30.8) | |
| CK5/6 (%) | ||||
| Negative | 819 (85.7) | 685 (84.6) | 134 (91.8) | 0.031 |
| Positive | 137 (14.3) | 125 (15.4) | 12 (8.2) | |
| Subtype (%) | ||||
| HR + HER2- | 664(69.5) | 561 (69.3) | 103 (70.5) | 0.031 |
| HR + HER2+ | 120(12.6) | 94 (11.6) | 26 (17.8) | |
| HR-HER2+ | 85(8.9) | 74 (9.1) | 11 (7.5) | |
| TNBC | 87(9.1) | 81 (10.0) | 6 (4.1) | |
| TNM (%) | ||||
| T1 | 407 (42.6) | 343 (42.3) | 64 (43.8) | 0.807 |
| T2 | 549 (57.4) | 467 (57.7) | 82 (56.2) | |
| Nipple and skin invasion (%) | ||||
| NO | 915 (95.7) | 779 (96.2) | 136 (93.2) | 0.151 |
| YES | 41 (4.3) | 31 (3.8) | 10 (6.8) | |
| Pathological type (%) | ||||
| NST | 923 (96.5) | 785 (96.9) | 138 (94.5) | 0.031 |
| ILC | 33 (3.5) | 25(3.1) | 8 (5.5) | |
| Multicenter (%) | ||||
| NO | 862 (90.2) | 733 (90.5) | 129 (88.4) | 0.517 |
| YES | 94 (9.8) | 77 (9.5) | 17 (11.6) | |
| Histological grade (%) | ||||
| I | 12 (1.3) | 10 (1.2) | 2 (1.4) | 0.917 |
| II | 484 (50.6) | 408 (50.4) | 76 (52.1) | |
| III | 460 (48.1) | 392 (48.4) | 68 (46.6) | |
| Lymph vascular invasion (%) | ||||
| NO | 612 (64.0) | 565 (69.8) | 47 (32.2) | < 0.001 |
| YES | 344 (36.0) | 245 (30.2) | 99 (67.8) | |
| Neurological invasion (%) | ||||
| NO | 825 (86.3) | 732 (90.4) | 93 (63.7) | < 0.001 |
| YES | 131 (13.7) | 78 (9.6) | 53 (36.3) | |
| Microcalcification (%) | ||||
| NO | 795 (83.2) | 673 (83.1) | 122 (83.6) | 0.983 |
| YES | 161 (16.8) | 137 (16.9) | 24 (16.4) | |
| Positive SLN, median (IQR) | 1.00 [1.00, 2.00] | 1.00 [1.00, 2.00] | 2.00 [1.00, 3.00] | < 0.001 |
| Negative SLN, median (IQR) | 3.00 [1.00, 4.00] | 3.00 [2.00, 4.00] | 1.50 [1.00, 3.00] | < 0.001 |
| Tumor size, median (IQR) | 22.00[17.00,27.00] | 22.00[17.00,26.00] | 22.00[18.00,30.00] | 0.144 |
ER Estrogen receptor, PR Progesterone receptor, HER2 Human epidermal growth factor receptor 2, TNBC Triple-Negative Breast Cancer, SLN Sentinel Lymph Node, UOQ Upper Outer Quadrant, NST No Special Type, ILC Invasive Lobular Carcinoma, HR + HER2- Luminal A, HR + HER2+ Luminal B, HR-HER2+ HER-2 overexpression
Figure 2, the Boruta algorithm retained 14 important variables after iterative importance assessment, while LASSO regression selected 11 key variables based on the λ1se criterion. Finally, combined with clinical, seven key predictors were identified: Neurological invasion、Lymph vascular invasion、Imaging abnormal nodes、Lymph nodes size、Tumor size、Negative SLN、Positive SLN.
Fig. 2.
The process of selecting variables by LASSO regression and Boruta algorithm. A LASSO coefficient profiles for 24 variables (plotted against log(λ) sequence); the optimal penalty coefficient λ was selected using 10-fold cross-validation and minimization criteria. B For LASSO model optimization, a binomial deviance curve was plotted against log(λ) (with dotted vertical lines), and 4 variables with non-zero coefficients were selected via λ₁ₛₑ. C The Boruta algorithm retained 14 important variables after iterative feature sorting
Model development and comparison
Nine machine learning models were selected for high nodal burden prediction: Random Forest (RF), Logistic Regression (LR), Gradient Boosting Machine (GBM), Neural Network (NN), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM), and Support Vector Machine (SVM). ROC, calibration curve, Decision Curve and confusion matrix of each model are shown in Figs. 3 and 4. Table 2 and Table 3 describes the performance metrics of the nine models in the training and test sets. Evaluation indicators include F1 score, Sensitivity, Specificity, Area Under Curve (AUC) and Accuracy. On the training set, all models achieved strong discrimination: Random Forest (AUC 0.971, 95% CI 0.948–0.994) and K-Nearest Neighbors (KNN; AUC 0.976, 95% CI 0.967–0.986) were nearly “perfect,” while the remaining algorithms delivered AUCs between 0.80 and 0.95—most notably Gradient Boosting Machine (GBM; AUC 0.932, 95% CI 0.910–0.955) and LightGBM (AUC 0.952, 95% CI 0.933–0.972). Nevertheless, several tree-based models, including Random Forest and LightGBM, exhibited significant performance deterioration in the validation cohort relative to the training set, suggesting a potential overfitting risk despite their excellent apparent discriminative ability. However, validation-set performance revealed clear overfitting: GBM and LightGBM AUCs fell to 0.785 and 0.756, respectively, and Random Forest and KNN dropped to 0.778 and 0.765, indicating limited generalizability. By contrast, Logistic Regression (training → validation AUC: 0.855 → 0.850; accuracy: 0.758 → 0.857; F1-score: 0.515 → 0.602) and Support Vector Machine (SVM; training → validation AUC: 0.853 → 0.857; accuracy: 0.758 → 0.853; F1-score: 0.512 → 0.596) exhibited the smallest performance gaps and achieved a balanced sensitivity (≈ 0.72) and specificity (≈ 0.88) on the validation set. Given their stable generalization and inherent interpretability, we selected Logistic Regression and SVM as the final candidates for feature‐importance and decision‐threshold analyses. According to the DeLong test, these AUC values were statistically significant (P < 0.005) (Table S2). Based on ROC curve analysis, the optimal probability cut-off for predicting high nodal burden was determined using the Youden index. A threshold of 0.45 provided the best balance between sensitivity (0.72) and specificity (0.88) and was therefore adopted for subsequent clinical interpretation.
Fig. 3.
Confusion Matrices of ML models. A LightGBM’s Confusion Matrices. B AdaBoost’s Confusion Matrices. C CKNN’s Confusion Matrices. D Logistic’s Confusion Matrices. E SVM’s Confusion Matrices. F GBM’s Confusion Matrices. G XGBoost’s Confusion Matrices. H Random Forest’s Confusion Matrices. I Neural Network’s Confusion Matrices
Fig. 4.
ROC curves, calibration curves and decision curve analysis of nine models. ROC Curves for Nine Machine Learning Models: A Training Set, B Test Set. Decision Curves for Nine Machine Learning Models: E Training Set, F Test Set. Calibration Curves for Nine Machine Learning Models: C Training Set, D Test Set
Table 2.
Predictive performance metrics of different machine learning algorithms of the test set
| Logistic | AUC | F1 score | SEN | SPE | Accuracy |
|---|---|---|---|---|---|
| 0.850 (0.779–0.920) | 0.602 | 0.721 (0.579–0.861) | 0.881(0.839–0.920) | 0.857 (0.815–0.899) | |
| SVM | 0.857 (0.790–0.923) | 0.596 | 0.721 (0.579–0.846) | 0.877(0.831–0.916) | 0.853 (0.811–0.895) |
| GBM | 0.785 (0.696–0.875) | 0.542 | 0.674 (0.537–0.805) | 0.856 (0.810–0.898) | 0.829 (0.787–0.871) |
| NN | 0.820 (0.739–0.900.739.900) | 0.614 | 0.721 (0.583–0.857) | 0.889 (0.851–0.927) | 0.864 (0.825–0.906) |
| RF | 0.778 (0.697–0.858) | 0.504 | 0.674 (0.520–0.810) | 0.823 (0.776–0.867) | 0.801 (0.755–0.843) |
| Xgboost | 0.835 (0.761–0.909) | 0.609 | 0.651 (0.500–0.791.500.791) | 0.914 (0.876–0.946) | 0.874 (0.832–0.913) |
| KNN | 0.765 (0.673–0.858) | 0.514 | 0.651 (0.500–0.795.500.795) | 0.844 (0.797–0.887) | 0.815 (0.769–0.857) |
| Adaboost | 0.742 (0.662–0.821) | 0.533 | 0.558 (0.412–0.705) | 0.905 (0.867–0.939) | 0.853 (0.811–0.892) |
| LightGBM | 0.756 (0.666–0.847) | 0.536 | 0.605 (0.447–0.744) | 0.885 (0.842–0.924) | 0.843 (0.797–0.885) |
Table 3.
Predictive performance metrics of different machine learning algorithms of the training set
| Logistic | AUC | F1 score | SEN | SPE | Accuracy |
|---|---|---|---|---|---|
| 0.855 (0.815–0.895) | 0.515 | 0.835 (0.762–0.907) | 0.744 (0.710–0.780) | 0.758 (0.727–0.791) | |
| SVM | 0.853 (0.814–0.893) | 0.512 | 0.825 (0.750–0.897) | 0.746 (0.710–0.779) | 0.758 (0.725–0.790) |
| GBM | 0.932 (0.910–0.955) | 0.601 | 0.942 (0.892–0.982) | 0.783 (0.750–0.815) | 0.807 (0.779–0.836) |
| NN | 0.880 (0.846–0.914) | 0.530 | 0.864 (0.795–0.926) | 0.746 (0.710–0.784) | 0.764 (0.733–0.797) |
| RF | 0.971 (0.948–0.994) | 0.970 | 0.942 (0.890–0.981) | 1.000 (1.000–1.000) | 0.991 (0.984–0.997) |
| Xgboost | 0.881 (0.843–0.918) | 0.595 | 0.757 (0.680–0.833) | 0.857 (0.827–0.886) | 0.842 (0.816–0.870) |
| KNN | 0.976 (0.967–0.986) | 0.824 | 0.981 (0.951–1.000.951.000) | 0.928 (0.906–0.948) | 0.936 (0.916–0.954) |
| Adaboost | 0.724 (0.674–0.775) | 0.529 | 0.524 (0.431–0.615) | 0.917 (0.894–0.939) | 0.857 (0.828–0.884) |
| LightGBM | 0.952 (0.933–0.972) | 0.690 | 0.854 (0.783–0.921) | 0.887 (0.859–0.912) | 0.882 (0.855–0.904) |
AUC area under the receiver operating characteristic curve, F1 Score harmonic mean of precision and recall, SEN sensitivity, SPE specificity;
Model explainability
The ultimate goal of constructing a predictive model is to inform clinical decision-making. Accordingly, we translated our logistic regression (LR) model into a user-friendly nomogram. Four variables dominate the scoring scale: sentinel lymph node (SLN) positivity, SLN negativity, lymphovascular invasion, and neurological invasion. In the nomogram, each additional positive SLN confers roughly 50 points (0–100 points for 1–3 positive SLNs), whereas each additional negative SLN adds approximately 12.5 points (0–100 points for 0–8 negative SLNs). A positive lymphovascular invasion status contributes about 50 points over negative, and similarly, positive neurological invasion contributes about 50 points over negative. Clinicians align a patient’s individual values on these axes to derive sub-scores (each ranging from 0 to 100), then sum them to yield a total score between 0 and 300. This total maps onto the linear predictor (–6 to 3), which in turn corresponds to a predicted probability of high nodal burden (0.1 to 0.9). Figure 5A and B illustrate this process in two exemplar patients: Patient 1 accrued a total of 269 points, corresponding to a 0.0808 probability of high nodal burden, whereas Patient 2 totaled 384 points, corresponding to a 0.87 probability.
Fig. 5.
Nomogram outputs of the logistic-regression model for two representative patients: A one patient predicted to have low nodal burden; B one patient predicted to have high nodal burden
For the more complex support vector machine (SVM) model, we employed SHapley Additive exPlanations (SHAP) to elucidate each predictor’s effect on high lymph node burden. In the SHAP summary plot (Fig. 6C), neurological invasion emerges as the single strongest driver, followed— in descending order—by SLN positivity, SLN negativity, lymphovascular invasion, lymph node size, Imaging abnormal nodes, and tumor size. Each point’s hue, from purple (lower SHAP value) to yellow (higher SHAP value), indicates how powerfully that feature pushes the model’s probability estimate upward or downward. The corresponding beeswarm plot (Fig. 6D) overlays feature value distributions to show how high versus low measurements stratify risk. Partial dependence plots for the top four predictors (Fig. 6E) further illuminate how pairwise interactions modulate predicted probabilities. Finally, Fig. 6F contrasts SHAP contribution stacks for one high-burden and one low-burden patient, demonstrating on a per-case basis how individual risk factors combine to yield an interpretable, patient-specific prediction.
Fig. 6.
SHAP analysis and interpretation. A SHAP beeswarm plot showing the distribution of SHAP values for each feature across all samples. Each dot represents an individual sample; colors indicate feature value (yellow = high, purple = low), and the x-axis represents SHAP value. B Ranking of feature importance based on the mean absolute SHAP values across all samples. Features with larger values have a greater average impact on model predictions. C SHAP dependency plots for the four most important features. D SHAP force plot of two patients
BC subtype analysis
Further analyses were performed to establish a high nodal burden prediction model according to BC subtype. The cohort for this analysis consisted of 664 (69.5%) patients with HR+/HER2- BC, 120 (12.6%) with HR+/HER2 + BC, 85 (8.9%) with HR-/HER2 + BC, and 87 (9.1%) with TNBC. The AUC, accuracy, sensitivity, and specificity of the high nodal burden prediction models according to BC subtype are described in Table S4. Among the four subtypes, TNBCs demonstrated the highest predictive performance, with a Logistic model AUC of 0.947 (95% CI: 0.861–1.000.861.000), followed by the HR-/HER2+ (SVM AUC: 0.905) and HR+/HER2+ (Logistic AUC: 0.889)) subtypes (Fig. 7). In contrast, the HR+/HER2- subtype had the lowest AUC (0.803 for Logistic). Model performance exhibited subtype-specific characteristics. For instance, in the HR+/HER2 + subgroup, the SVM model achieved exceptionally high sensitivity (0.962) but lower specificity, whereas the Logistic model provided a more balanced performance (F1-score: 0.712). Conversely, predictive models for both TNBC and HR-/HER2 + subtypes showed robust and balanced performance across all metrics (AUCs > 0.89).
Fig. 7.
ML model for high nodal burden prediction according to breast cancer subtype: A Triple negative breast cancer, B HR+/HER2-, C HR + HER2+, D HR-HER2+
Discussion
In recent years, stepwise de-escalation of axillary management has become a major focus in breast cancer research. Landmark trials such as ACOSOG Z0011, AMAROS, and SINODAR-ONE have consistently shown that omitting axillary lymph node dissection (ALND) in clinically node-negative (cN0) patients with 1–2 positive sentinel lymph nodes (SLNs) does not compromise long-term survival and significantly reduces complications like lymphedema [19]. Further analyses, however, reveal considerable residual nodal disease in certain patients. For instance, 13.7% of patients in ACOSOG Z0011, 8.0% in AMAROS, and 22.1% in the OTOASOR trial had ≥ 4 metastatic nodes or were classified as ≥ pN2a. Despite this, omission of completion ALND (CALND) did not increase long-term regional recurrence. These findings align with the 15.3% prevalence of high nodal burden (≥ pN2) observed in our study, highlighting a clinical dilemma: avoiding overtreatment with ALND while accurately identifying high-risk patients to optimize adjuvant therapy [20].
We included patients with 1–3 SLN macrometastases to enhance real-world applicability. Although patients with three positive SLNs are typically advised to undergo ALND, data on ALND omission in this subgroup remain limited. Their inclusion improves the clinical relevance of our study and aligns with de-escalation efforts extending beyond the conventional 1–2 SLN threshold. To address existing contradictions, we developed a machine learning-based model aimed at more accurately identifying patients with high nodal burden. This approach is especially valuable when ALND is omitted, as it helps bridge the staging information gap by detecting individuals with only 1–3 positive SLNs who still carry a significant risk of ≥ pN2 disease. Such a tool is critical for personalizing adjuvant systemic and radiation therapy beyond the binary ALND decision. We systematically evaluated nine algorithms, among which logistic regression (LR) and support vector machine (SVM) achieved the most favorable balance between generalizability and predictive performance. Feature selection was conducted using both LASSO regression and the Boruta algorithm. LASSO employs L1 regularization to enhance sparsity by retaining the most contributive variables, thereby improving generalizability and accuracy. Its hybrid design offers robustness and interpretability, making it suitable for high-dimensional biomedical data. In contrast, Boruta iteratively refines selection thresholds via random shadow features, capturing predictors with potential biological relevance even when statistically subtle [21] By integrating both methods with clinical expertise, we established the final feature set, reinforcing the scientific rigor of variable selection.
As a classical clinical prediction approach, logistic regression exhibited notable stability, with minimal AUC difference between training and validation cohorts (0.855 vs. 0.850), and improved accuracy (0.758 to 0.857) and F1 score (0.515 to 0.602). This consistency underscores the model’s robustness and adaptability across sample distributions. When translated into a nomogram, it further offers an intuitive tool to support clinical decision-making. In contrast, the SVM model demonstrated superior performance optimization, achieving a slightly higher validation AUC of 0.857 and the best balance between specificity (0.88) and sensitivity (0.72). This strength stems from SVM’s capacity to handle complex nonlinear relationships via kernel functions [22] Notably, SVM excelled in capturing interactions between key variables—such as the number of positive SLNs and lymphovascular invasion—outperforming logistic regression by approximately 8% in identifying high-burden risk in cases with ≥ 2 positive SLNs and lymphovascular invasion. Although the “black box” nature of SVM has historically hindered clinical adoption [23], we applied SHAP analysis in this study to visualize the model’s decision logic, providing interpretable support for its potential translation into practice.
Feature selection represents a critical step in model development, directly influencing predictive performance. In this study, SHAP was employed to quantify each feature’s contribution to the model’s predictions. SHAP values enabled both global and local interpretability, revealing that neurological invasion was the dominant predictor of high nodal burden, with the highest mean SHAP value. Lymphovascular invasion also emerged as a strong determinant, underscoring the association between aggressive tumor biology and increased lymph node metastatic burden, in line with findings from prior studies [24]. The absolute number of axillary lymph nodes involved is considered the most important prognostic factor for breast cancer [25]. A greater number of metastatic sentinel lymph nodes (SLNs) generally indicates an increased likelihood of further lymph node involvement [26]. This, in turn, reflects an elevated risk of high nodal burden, consistent with the findings from SHAP analysis. However, the number of SLNs identified during surgery can be affected by factors such as surgeon experience, tracer selection, and patient-specific anatomical variations. In cases where too few SLNs are retrieved, the proportion of positive nodes may be artificially inflated, potentially overestimating the risk of non-sentinel lymph node (NSLN) metastasis. To mitigate this bias, the majority of SLNs detected in our cohort (897 patients, 93.8%) numbered two or more. Although pathological tumor size contributed only modestly to SHAP values, it is widely recognized as a predictor of lymph node metastasis and retains meaningful predictive value. Notably, when multiple risk factors such as lymph vascular invasion, neurological invasion, and radiologically abnormal lymph nodes coexist, our model yields a high predicted probability of ≥ 4 nodal metastases (e.g., ≥ 0.6–0.7). In these cases, ALND or enhanced nodal radiotherapy should be considered to prevent underestimation of axillary disease burden.
In addition to pathological parameters, the Boruta algorithm identified two lymph node–related features derived from preoperative ultrasound—namely, radiologically abnormal nodes and lymph node size. The assessment of radiologically abnormal nodes is based on the RI-RADS system, which serves as a critical reference for clinical lymph node staging and has been recognized as a predictive factor in previous similar modeling studies [27]. Lymph node size, although not included in previous predictive models, has been shown to change during lymph node metastasis in breast cancer [28]. This evidence provided a rationale for incorporating this variable into our model.
There has been a lot of research on models used to predict high nodal burden. In the multicenter risk prediction tool for ≥ 4 axillary lymph node (ALN) metastases in SLN-positive breast cancer developed by Meretoja et al. [29], based on retrospective data from 675 patients, the model achieved an AUC of 0.77 (95% CI, 0.74–0.81) in external validation using a separate retrospective cohort of 760 patients. The variables incorporated into the model included the prevalence of ≥ pN2 by tumor site, the number of metastatic and benign SLNs, tumor size, and the presence of extracapsular extension. In a nomogram developed by Yang et al. [10], which included the number of SLN metastases, extracapsular extension, pT stage, and tumor location within the breast for 1,480 patients, the model achieved an AUC of 0.80 (95% CI, 0.76–0.85) in the external validation cohort. In a smaller study by Corsi et al. (n = 271) [30], the SLN ratio, extranodal extension, and multifocal disease were identified as predictors of more than two metastatic axillary lymph nodes, yielding an AUC of 0.74. No confidence interval was reported, and the model was not externally validated. In contrast, the present study not only incorporated conventional postoperative pathological data but also integrated preoperative ultrasound information of lymph nodes, thereby leveraging multidimensional data for model construction. Moreover, beyond employing classical logistic regression, we implemented and systematically compared several machines learning algorithms, representing an innovative application of machine learning techniques in the prediction of high nodal burden.
This study has several limitations. Most importantly, the absence of external validation represents a major limitation, as the model was developed and tested using data from a single center with a retrospective design, which may introduce selection bias and limit its generalizability. Second, the model was developed exclusively in SLNB-positive patients, focusing on the clinical scenario in which the omission of ALND is most controversial; however, this inevitably restricts its applicability to a broader breast cancer population. Finally, although key clinicopathological variables were incorporated, other potentially informative predictors, such as genomic assay results, were not available for analysis.
Conclusion
In this study, we developed and internally validated machine-learning models to estimate the risk of high axillary nodal burden in breast cancer patients with positive sentinel lymph nodes. Both the logistic regression and support vector machine models demonstrated stable discrimination and calibration across datasets. By integrating perioperative clinicopathological variables, these models provide an interpretable and practical approach to postoperative risk stratification. Rather than replacing clinical judgment, such tools may assist clinicians in identifying patients at higher risk of extensive nodal disease when axillary lymph node dissection is omitted. External validation and prospective multicenter studies are required before these models can be considered for routine clinical application.
Supplementary Information
Authors’ contributions
S.K.L and T.S.X conceived and planned the experiment, and X.Y.Y and Y.F.J contributed to the collection and collation of clinical data. H.X.L contributed to the interpretation of the results. S.K.L wrote this manuscript. T.S.X oversaw the project. All authors reviewed and confirmed the manuscript.
Funding
Funding was supported by Program of Jiangsu Province Hospital (KF24).
Data availability
The datasets used and analyzed during the current study available from the corresponding author on reasonable request.
Declarations
Ethics approval and consent to participate
Ethical approval for this retrospective study was obtained from the Ethics Committee of The First Affiliated Hospital of Nanjing Medical University (2025-SR-653). In compliance with institutional and regulatory requirements, the need for obtaining informed consent was formally waived by the committee. The research was conducted in accordance with the ethical standards set forth in the declaration.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
ShuaiKang Li and Xingying Yu contributed equally to this work.
References
- 1.Sávolt Á, Peley G, Csaba P, et al. Eight-year follow up result of the OTOASOR trial: the optimal treatment of the Axilla - Surgery or radiotherapy after positive Sentinel lymph node biopsy in early-stage breast cancer: A randomized, single centre, phase III, non-inferiority trial. Eur J Surg Oncol. 2017;43(4):672–9. 10.1016/j.ejso.2016.12.011. [DOI] [PubMed] [Google Scholar]
- 2.Kim J, Harper A, McCormack V, et al. Global patterns and trends in breast cancer incidence and mortality across 185 countries. Nat Med. 2025;31(4):1154–62. 10.1038/s41591-025-03292-2. [DOI] [PubMed] [Google Scholar]
- 3.Scarping I, Bendahl PO, Szulkin R, et al. Prediction of high nodal burden in patients with Sentinel Node–Positive luminal ERBB2-Negative breast cancer. JAMA Surg. 2024;159(12):1393–401. 10.1001/jamasurg.2024.3036. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Giuliano AE, Ballman KV, McCall L, et al. Effect of axillary dissection vs no axillary dissection on 10-Year overall survival among women with invasive breast cancer and Sentinel node metastasis: the ACOSOG Z0011 (Alliance) randomized clinical trial. JAMA. 2017;318(10):918–26. 10.1001/jama.2017.11470. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Galimberti V, Cole BF, Zurrida S, et al. Axillary dissection versus no axillary dissection in patients with sentinel-node micrometastases (IBCSG 23 – 01): a phase 3 randomised controlled trial. Lancet Oncol. 2013;14(4):297–305. 10.1016/S1470-2045(13)70035-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Murata T, Watase C, Shiino S, et al. Development and validation of a pre- and intra-operative scoring system that distinguishes between non-advanced and advanced axillary lymph node metastasis in breast cancer with positive Sentinel lymph nodes: a retrospective study. World J Surg Oncol. 2022;20(1):2. 10.1186/s12957-021-02473-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.EBCTCG (Early Breast Cancer Trialists’ Collaborative Group), McGale P, Taylor C, et al. Effect of radiotherapy after mastectomy and axillary surgery on 10-year recurrence and 20-year breast cancer mortality: meta-analysis of individual patient data for 8135 women in 22 randomised trials. Lancet. 2014;383(9935):2127–35. 10.1016/S0140-6736(14)60488-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Rastogi P, O’Shaughnessy J, Martin M, et al. Adjuvant abemaciclib plus endocrine therapy for hormone receptor–Positive, human epidermal growth factor receptor 2–Negative, High-Risk early breast cancer: results from a Preplanned monarche overall survival interim Analysis, including 5-Year efficacy outcomes. J Clin Oncol. 2024;42(9):987–93. 10.1200/JCO.23.02192. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Cardoso F, van’t Veer LJ, Bogaerts J, et al. 70-Gene signature as an aid to treatment decisions in Early-Stage breast cancer. N Engl J Med. 2016;375(8):717–29. 10.1056/NEJMoa1602253. [DOI] [PubMed] [Google Scholar]
- 10.Yang Z, Lan X, Huang Z, et al. Development and external validation of a nomogram to predict four or more positive nodes in breast cancer patients with one to three positive Sentinel lymph nodes. Breast. 2020;53:143–51. 10.1016/j.breast.2020.07.009. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Park YH, Karantza V, Calhoun SR, et al. Prevalence, treatment patterns, and prognosis of low Estrogen receptor-positive (1% to 10%) breast cancer: a single institution’s experience in Korea. Breast Cancer Res Treat. 2021;189(3):653–63. 10.1007/s10549-021-06309-1. [DOI] [PubMed] [Google Scholar]
- 12.Wolff AC, Hammond MEH, Hicks DG, et al. Recommendations for human epidermal growth factor receptor 2 testing in breast cancer: American society of clinical Oncology/College of American pathologists clinical practice guideline update. J Clin Oncol. 2013;31(31):3997–4013. 10.1200/JCO.2013.50.9984. [DOI] [PubMed] [Google Scholar]
- 13.Ignatiadis M, Buyse M, Sotiriou C. St Gallen international expert consensus on the primary therapy of early breast cancer: an invaluable tool for physicians and scientists. Ann Oncol. 2015;26(8):1519–20. 10.1093/annonc/mdv259. [DOI] [PubMed] [Google Scholar]
- 14.Gucalp A, Tolaney S, Isakoff SJ, et al. Phase II trial of bicalutamide in patients with androgen receptor-positive, Estrogen receptor-negative metastatic breast cancer. Clin Cancer Res. 2013;19(19):5505–12. 10.1158/1078-0432.CCR-12-3327. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Gown AM. Current issues in ER and HER2 testing by IHC in breast cancer. Mod Pathol. 2008;21(Suppl 2):S8–15. 10.1038/modpathol.2008.34. [DOI] [PubMed] [Google Scholar]
- 16.Wolff AC, Hammond MEH, Allison KH, et al. Human epidermal growth factor receptor 2 testing in breast cancer: American society of clinical Oncology/College of American pathologists clinical practice guideline focused update. J Clin Oncol. 2018;36(20):2105–22. 10.1200/JCO.2018.77.8738. [DOI] [PubMed] [Google Scholar]
- 17.Hehman G, Koch HK, Rah CH, et al. Peering into the black box of artificial intelligence: evaluation metrics of machine learning methods. AJR Am J Roentgenol. 2019;212(1):38–43. 10.2214/AJR.18.19996. [DOI] [PubMed] [Google Scholar]
- 18.Lundberg SM, Nair B, Vavilala MS, et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat Biomed Eng. 2018;2(10):749–60. 10.1038/s41551-018-0304-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Tinterri C, Gentile D, Gatzemeier W, et al. Preservation of axillary lymph nodes compared with complete dissection in T1-2 breast cancer patients presenting one or two metastatic Sentinel lymph nodes: the SINODAR-ONE multicenter randomized clinical trial. Ann Surg Oncol. 2022;29(9):5732–44. 10.1245/s10434-022-11838-0. [DOI] [PubMed] [Google Scholar]
- 20.Kaur AM, Fairchild L, Glaser CL, et al. Partial-breast radiotherapy after breast conservation surgery for women with early breast cancer (UK IMPORT LOW): 10-year outcomes from a multicentre, open-label, randomised, controlled, phase 3, non-inferiority trial. Lancet Oncol. 2025;26(7):898–910. 10.1016/S1470-2045(25)00367-9. [DOI] [PubMed] [Google Scholar]
- 21.Díaz-Uriarte R, Alvarez de Andrés S. Gene selection and classification of microarray data using random forest. BMC Bioinformatics. 2006;7:3. 10.1186/1471-2105-7-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Bhardwaj A, Jain A. A systematic review of machine learning algorithms for breast cancer detection. Tissue Cell. 2025;95:102929. 10.1016/j.tice.2025.102929. [DOI] [PubMed] [Google Scholar]
- 23.Kumar A, Agarwal M, Singh R, et al. Support matrix machine: A review. Neural Netw. 2025;181:106767. 10.1016/j.neunet.2024.106767. [DOI] [PubMed] [Google Scholar]
- 24.Chen H, Mao X, Ma X, et al. Correlation analysis of pathological features and axillary lymph node metastasis in patients with invasive breast cancer. J Immunol Res. 2022;2022:7150304. 10.1155/2022/7150304. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.de Boer M, van Deurzen CHM, van Dijck JAAM, et al. Micrometastases or isolated tumor cells and the outcome of breast cancer. N Engl J Med. 2009;361(7):653–63. 10.1056/NEJMoa0904832. [DOI] [PubMed] [Google Scholar]
- 26.Liu H, Yang P, Guo C, et al. Development and validation of nomograms for predicting axillary non-SLN metastases in breast cancer patients: A retrospective analysis. Front Oncol. 2023;13:1096589. 10.3389/fonc.2023.1096589. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Yang L, Zhao X, Yang L, et al. A new prediction nomogram of non-sentinel lymph node metastasis in cT1-2 breast cancer patients with positive Sentinel lymph nodes. Sci Rep. 2024;14(1):12976. 10.1038/s41598-024-63855-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Ma X, Tang X, Pan H, et al. B cells and IL-21-Producing follicular helper T cells cooperate to determine the dynamic alterations of premetastatic tumor draining lymph nodes of breast Cancer. Res (Wash D C). 2024;7:0346. 10.34133/research.0346. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Meretoja TJ, Aud RA, Heikkila PS, et al. International multicenter tool to predict the risk of four or more tumor-positive axillary lymph nodes in breast cancer patients with Sentinel node macrometastases. Breast Cancer Res Treat. 2013;138(3):817–27. 10.1007/s10549-013-2492-3. [DOI] [PubMed] [Google Scholar]
- 30.Corsi F, Sorrentino L, Sartani A, et al. Prediction of nodal staging in breast cancer patients with 1–2 Sentinel nodes in the Z0011 era. Med (Baltim). 2020;99(24):e21721. 10.1097/MD.0000000000021721. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets used and analyzed during the current study available from the corresponding author on reasonable request.